We have an issue with unstable link on an AGX Xavier, we think we narrowed it down to broadcast/multicast, in our network it can happen at times that a couple dozen multicasts are received in a ~50-100ms range, we now see that this causes the ethernet controller to go down and come back up without so much as an error.
We see that putting the controller in promiscuous mode by doing a tcpdump, the issue disappears, and that seems to bypass something for the L2 receiving.
So our question is how we can ensure multicast/broadcast doesn’t bring down the controller or how we can disable this functionality (we’re not sure what is even enabling this) in the device tree
I won’t be able to help on this, but to see what is going on I recommend you get a full serial console boot log, and don’t stop logging once fully booted and logged in. Then run “dmesg --follow” in the serial console. Do whatever you need to get the system to fail. The log should capture the failure. Without that it’d be rather difficult to even guess.
We’ve tried everything in that regard, I’ve enabled debug in all the seemingly relevant drivers, nvethernet, our phy, nothing is really showing much, we’re still trying to figure out a proper cause for the link down but there’s not so much as an indication as to why
There is an issue with the TI DP83867 phy we used where if pin strapping for a certain pin that controls autoneg is not done, as was not the case on our board, the phy exposes an Energy Efficient Ethernet register that claims to support EEE on 100/1000Mbps
This gets incorrectly (or correctly because that’s what it reads) picked up by the Nvidia ethernet controller who then enables EEE and Tx LPI on it’s MAC, connected to a switch supporting this feature was what then caused this very strange behavior